Accounted's Three-Tier Reasoning Engine Explained

Why Simple Rules Are Not Enough

Most bookkeeping software categorises transactions using simple rules. If the merchant name contains "SCREWFIX," categorise as "Stock and materials." If it contains "SHELL," categorise as "Travel costs." These rules work for straightforward, recurring transactions -- your monthly phone bill, your regular fuel purchases, your subscription to Microsoft 365.

But rules break down quickly in the real world. What about a payment to "Amazon" that could be office supplies, stock, or personal? What about a cafe transaction that might be a business meeting or just a coffee? What about a payment to someone's name, which could be a subcontractor, a professional service, or something entirely personal?

Rules-based systems handle ambiguity by defaulting to the most common category and hoping for the best. That's fine 70-80% of the time. But 70-80% accuracy means one in four or five transactions might be wrong. For sole traders with 50 transactions a month, that's 10-15 potential errors -- every month.

Accounted's three-tier reasoning engine was built to solve this problem. Instead of relying on a single method, it uses three distinct tiers of analysis, each handling different types of transactions. The result is categorisation accuracy above 95%, with full transparency about confidence levels on every decision.

Tier 1: Deterministic Rules (95% of Transactions)

The first tier is the fastest and most reliable: deterministic rules that produce definitive answers for clear-cut transactions.

How Tier 1 Works

Tier 1 maintains a comprehensive database of UK merchants mapped to HMRC expense categories. It includes:

• Major retailer mappings -- Screwfix, Toolstation, Wickes, B&Q, Currys, and thousands more, each mapped to the appropriate category
• Utility and subscription providers -- EE, Vodafone, BT, Sky, Adobe, Microsoft, mapped to "Telephone" or "Office costs" as appropriate
• Fuel stations -- Shell, BP, Esso, Texaco, mapped to "Travel costs"
• Your personal rules -- as you confirm and correct categorisations, Tier 1 builds custom rules specific to your business

Tier 1 also handles amount-based rules (a £7.99 monthly charge to Netflix is clearly a subscription), frequency rules (payments that recur on the same date each month), and pattern rules (three consecutive transactions at a builder's merchant during working hours are almost certainly business expenses).

When Tier 1 Is Sufficient

Tier 1 handles approximately 95% of all transactions processed by Accounted. These are the ones where the merchant is recognisable, the pattern is clear, and the category is unambiguous. They're processed in milliseconds with near-perfect accuracy.

The key advantage of Tier 1 is speed and cost-efficiency. Deterministic rules require no AI model inference, which means they're instant and free to run. This keeps Accounted fast and affordable.

When Tier 1 Escalates

If Tier 1 can't produce a high-confidence answer -- because the merchant isn't recognised, the transaction is ambiguous, or the amount doesn't fit known patterns -- it passes the transaction to Tier 2.

Tier 2: Retrieval-Augmented Generation (4% of Transactions)

Tier 2 handles the transactions that rules can't resolve but that don't require a full language model. It uses a technique called retrieval-augmented generation (RAG) -- a method that combines database retrieval with AI reasoning.

How Tier 2 Works

When a transaction reaches Tier 2, the system:

1. Retrieves similar transactions from across the Accounted platform (anonymised -- no personal data is shared). If other sole traders in your industry have categorised payments to the same merchant, that information provides useful context.

2. Retrieves HMRC guidance relevant to the transaction. For example, if the merchant appears to be a training provider, Tier 2 pulls the relevant rules about whether training costs are allowable expenses.

3. Retrieves your historical patterns. Even if this specific merchant is new, the AI looks for analogous transactions in your history.

4. Combines these retrievals into a recommendation with a confidence score.

Practical Example

Say you make a payment of £150 to "Manchester Business School" -- a merchant Tier 1 doesn't recognise. Tier 2:

• Retrieves that other sole traders have categorised similar payments to "Business School" merchants as "Training costs"
• Retrieves HMRC guidance on training expenses (allowable if they update existing skills, not allowable if they teach entirely new skills)
• Notes that you've previously categorised other educational purchases as business expenses
• Produces a recommendation: "Training costs" with 82% confidence

Because the confidence is below 95%, Penny flags it for your confirmation via WhatsApp rather than categorising silently.

When Tier 2 Escalates

For the roughly 1% of transactions where even RAG can't produce a confident answer -- highly ambiguous cases, unusual merchants, or transactions that don't match any known pattern -- the system escalates to Tier 3.

Tier 3: Large Language Model Reasoning (1% of Transactions)

Tier 3 brings in the full power of a large language model for the most complex categorisation decisions.

How Tier 3 Works

When a transaction reaches Tier 3, the LLM receives:

• The transaction details (merchant, amount, date, description)
• Your business type and industry
• Your recent transaction history for context
• Relevant HMRC rules and categories
• Any receipt data associated with the transaction

The LLM reasons about the transaction much like a human bookkeeper would, considering multiple factors simultaneously and weighing possibilities against each other.

Practical Example

A payment of £350 to "Green Valley Services" -- no merchant match, no similar transactions from other users, and the name is too generic for Tier 2 to resolve. Tier 3:

• Considers that "Green Valley" could be a landscaping company, a cleaning service, or a consultancy
• Notes that you're a property landlord
• Considers that £350 is a common amount for a one-off garden or property maintenance service
• Examines whether you've had similar payments for property maintenance
• Reasons that this is most likely property maintenance, but isn't certain
• Produces a recommendation at 65% confidence

At 65% confidence, Penny asks you directly:

Penny: I saw a payment of £350.00 to "Green Valley Services" on 24 Feb. I'm not sure about this one. Could you tell me what it was for?

Why Tier 3 Is Used Sparingly

LLM inference is slower and more expensive than rules or RAG. By handling 95% of transactions at Tier 1 and 4% at Tier 2, Accounted reserves the LLM for genuinely difficult cases where its reasoning capability adds real value.

This architecture means you get the speed and cost-efficiency of simple rules for straightforward transactions, the contextual intelligence of RAG for moderately complex ones, and the full reasoning power of an LLM for the genuinely ambiguous cases.

How the Tiers Work Together

The three tiers aren't separate systems -- they're a unified pipeline. Every transaction enters at Tier 1. If Tier 1 produces a confidence score above 95%, it's categorised and the process stops. If not, it flows to Tier 2. If Tier 2 can't reach 95% confidence either, Tier 3 takes over.

The result is a system that's:

• Fast -- 95% of transactions are categorised in milliseconds
• Accurate -- each tier applies the right level of analysis for the difficulty
• Transparent -- every categorisation includes a confidence score, regardless of which tier produced it
• Cost-effective -- expensive LLM calls are reserved for the small percentage of transactions that need them

This is what makes Accounted's categorisation fundamentally different from the rules-only approach used by traditional bookkeeping software. It's not just "AI" as a marketing term -- it's a carefully engineered system where each component handles what it does best.

The Learning Loop

All three tiers improve over time:

• Tier 1 gains new rules as you confirm categorisations. A merchant that started in Tier 2 or Tier 3 might become a Tier 1 rule after you've categorised it consistently a few times.
• Tier 2 benefits from aggregated (anonymised) categorisation data as more users join the platform. The retrieval database grows richer and more accurate.
• Tier 3 reasoning improves as the underlying language models are updated and fine-tuned with bookkeeping-specific knowledge.

The practical effect: Penny gets smarter the longer you use her, and the platform gets smarter the more users it serves. The ICAEW's research on AI in accounting identifies this kind of continuous learning as one of the most promising applications of machine learning in practice.

Read about how Penny categorises your transactions to see this in practice.

Why This Architecture Matters for You

You don't need to understand the technical details to benefit from the three-tier engine. What matters is the outcome:

• Your transactions are categorised accurately, with transparency about confidence
• The system is fast -- no waiting for "AI processing"
• It improves over time without you doing anything special
• When it's uncertain, it asks rather than guessing
• Your MTD submissions are based on accurate data
• Your real-time tax estimates are reliable

That's what the three-tier reasoning engine delivers. Not AI for the sake of AI, but a genuinely better way to keep your books accurate and your tax position clear.

Explore Accounted's features or start your free trial to experience the three-tier engine firsthand.

Useful Resources

• HMRC's MTD for Income Tax requirements
• HMRC's guide to allowable expenses